On the mechanism of hot deformation

To overcome the disadvantages of the phenomenological constitutive model, which is sensitive to data and limited by model structure and assumptions, and to enhance the prediction accuracy of the flow behavior of near-β titanium alloy during hot deformation, a machine learning prediction model was established using the whale optimized neural network algorithm (WOA-BP). To validate the model’s accuracy, hot compression experiments were conducted on a near-β titanium alloy, Ti-3Mo-6Cr-3Al-3Sn. Subsequently, the phenomenological constitutive model and WOA-BP model for the hot deformation process of the alloy were established. The analysis of flow stress prediction errors revealed significant improvements in comparison to the modified J-C constitutive structure model and Arrhenius constitutive structure model. Specifically, the WOA-BP model showed an increased error correlation coefficient (R) by 0.030063 and 0.17252, respectively, along with reduced average relative errors (AARE) to 14.92575 and 7.70414, respectively. The root mean square error (MSE) and mean absolute error (MAE) were significantly reduced to 22.51002 and 3.652993, respectively. The WOA-BP model greatly improved the accuracy of flow stress predictions. Using the flow stress prediction value from the WOA-BP model, the hot processing map was established at a true strain of 0.6. At a power dissipation factor (η) of 0.53–0.59, fully recrystallized grains appeared in the microstructure, exhibiting a relatively uniform grain size. Conversely, at η values of 0.17–0.21, significant deformation bands formed in the microstructure, making it unsuitable for thermal processing. This trend aligns with the power dissipation values, demonstrating the hot processing map’s accuracy.